Improving Web Vulnerability Scanning 1Daniel Zulla
IntroductionHey! 2 ■ Hi there! ■ I’m Dan. This is my ﬁrst year at DEFCON. ■ I do programming and security start-ups. ■ I do some penetration testing as well
More Introduction 3 ■ Today I’m going to talk about vulnerability scanning ■ Primary on the web ■ “The cloud” is involved as well ■ Network security too ■ I’ll show some things, so there is plenty of demo time ■ Have fun, thanks for being here!
Some Facts 4 ■ There are a lot of web vulnerability scanners, fuzzers and penetration testing tools out there already ■ Some of them work, some of them do not ■ But basically all of them have one thing in common: They actually don’t attack web applications on the application layer ■ They mostly fuzz HTTP and sometimes perform injection attacks
Some more facts 5 ■ The fundamental design of web scanners has not changed in over a decade ■ But: The web has changed. ■ So there seems to be a problem.
Software ArchitectureWhat web vulnerability scanners and fuzzers look like 6 Plugins A HTTP Library RXSS BSQLI EVAL The Core PXSS LFI OSC Multithreading / Output Engine Forking SQL RFI [...]
How I see it 9■ Both of them are right.■ The web is a mess. Nobody cares about RFCs anymore. (Especially these SEO guys!)■ 10 years ago, you would have expected a Query String at the end of a URL like https://foo.com/xxx/yyy?foo=bar■ Nowadays, https://foo.com/something.ext/foo/bar is good practice■ The result: It’s incredibly hard for scanner developers to ﬁgure out the dynamic components of an HTTP request. Because of that, we feel overhelmed and fuzz nearly everything.■ Header Keys, Header Values, VHost, Cookie, Method, Path, Version, ...
Web 2.0 11■ But - WE DO SECURITY■ Is it really our job to make sure that our software executed all the JS and grabbed all the links?■ When we spend 100 hours on the crawler, and 5 hours on the actual payloads (that’s how it looks right now) something, somewhere, went terribly wrong■ So - Is there a (open source?) piece of software that we could use instead of the HTTP library? Something that has prooven its mastery in handling unpredictably broken web content already? There is.
Software ArchitectureWhat it should look like 14 The Front-End A HTTP RXSS BSQLI EVAL The Core Library PXSS LFI OSC Reporting Engine SQL RFI [...] The Exploitation Engine
Changes? Improvments? 15 ■ Replacing the HTTP library by a Webkit Engine ■ Less code (A lot less code) ■ 100% support for JS/Ajax/Broken HTML/JS Events/Crazy Redirects and all kinds of things ■ The ability to simulate human user behaviour ■ CSS Renderings (Two text ﬁelds beside each other: 10px - one of them is a input[type=password]) - May be a login!
Making it scale (heavily) 16 ■ Webkit is slow (Website rendering, Executing JS, ... - compared to - Speaking Plaintext HTTP) ■ Downloading Images is slow ■ Waiting for delayed JS events is slow ■ Flash is even slower
Making it scale (heavily)Bad news: Qt / PyQt / PySide 17 ■ QtWebkit does not support multithreading ■ It tends to SEGFAULT from time to time :( ■ Multiple QApplication instances are almost impossible to handle in one Python namespace
Making it scale (heavily)Good news: Building a preforking TCP Server 18 ■ Spawning a pool of processes works quite well (one QApplication +one Browser instance per Process) ■ Simultaneous downloads ■ Better accessibility inside the scanner (multiprocessing insides loops to increase performance)
Mastering Authentication 20 ■ There is no such thing as a standarized web login ■ Basically, everybody develops access control on the web slightly differently ■ You can try to detect them by the name/id of the attributes, but that is not reliable ■ But in the end, Web logins generally have a few things in common that makes them easily detectable. At least, for our browser engine
Mastering AuthenticationNot more than 2 visible (!) text ﬁelds 21 has_login_texts()
Mastering Authentication Geometry! Usually, the two visible text ﬁelds are under(), next_to() or at least near(radius=10px) each other 23X1 = X2 X1 = X2 ! Y1 = Y2
Mastering Authentication 24 ■ That was easy! ■ The common way to solve that problem, is to iterate through a wordlist (login, auth, signin, [...]) while checking the input[id], input[name] attributes ■ That’s not necessarily wrong or bad practice ■ After putting the pieces together: ■ .login(“username”, “password”)
Mastering AuthenticationWhen we are signed in 26 ■ New problems occur: How can we let the scanner check if we are indeed signed in? ■ Common practive: Looking for a /logout/i String ■ The problem: Inefﬁcient. Likely to cause false positives ■ There has to be a better way: ■ Introduction “Strategies”
Strategy.AuthenticationStep 1: Identiﬁcation 27 ■ Identifying a login form (3-way approach, input[type=password], geometry, [...])
Strategy.AuthenticationStep 3: Going in: .login(“..”, “..”) 29 ■ Verifying valid credentials - Behaviour should not be similiar to the behaviour of a invalid login
Strategy.AuthenticationStep 4: Going out. .logout() 30 ■ Doing similiar work again for .logout() function seems obsolote ■ But it really isn’t. ■ It is the basis to a .is_still_loggedin() function ■ Which is really important to stay logged in during crawling ■ And if the scanner logged itself out, it can simply .login() again ■ That’s cool. :-)
Exploitation and Privilege Escalation 31■ There is a whole universe besides injection vulnerabilities■ Usually, scanners don’t detect them■ But they should■ And now they can: .login(“user1”, “...”); .logout(); .login(“user2”, “...”)■ => Demo Time: Privilege Escalation, Multi-User Systems
Geographically distributed scanning:Using the cloud 32 ■ When (injection) vulnerabilities are getting complicated: ■ Scenario 1: The backend of a website creates a log entry for every new IP address. It logs the USERAGENT. The log entries are kept in a SQL database. The function that creates the log entries, is vulnerable. The User-Agent is injectable. The problem is: ■ It only works once. As soon as the IP is in the database, the function won’t be executed anymore :-( ■ ==> SQLMap (and every other tool) will fail.
Geographically distributed scanning:Using the cloud 33 ■ But they shouldn’t! ■ The limitation is totally detectable ■ And a new IP is just as far away as a single EC2 API call
Geographically distributed scanning:Using the cloud 34 ■ Indeed! The cloud is a good thing for security :) ■ Demo Time: Introducing: sqlmap and w3af (on steroids)
Combining “Strategies” and thedistributed scanning 35 ■ Introducing next generation vulnerability scanning ■ Exploiting a really amazingly hard SQL Injection ■ Demo Time
Further Research & Additional Ideas 36 ■ Country speciﬁc restrictions can be by-passed in a fully automatic manner ■ (Error) messages can be parsed and interpreted: Wolfram Alpha ■ Bloomﬁlters should be integrated ■ Other “Strategies” should be implemented (the limitations are gone)
More Live Demos 37 ■ Demonstrating a logical layer beyond Authentication: .pay(“0000111122223333”, CVV=121, type=VISA) .search(“search query”) .sort(“DESC UNION SELECT [...]”) ■ Interpreting error messages ■ Pivoting on penetrated hosts - Spawning another scanner instance ■ And ﬁnally: Reporting!